AI and Creativity — Can Machines Be Genuinely Creative?
The question splits into two separable problems: (1) Can AI produce creative outputs? and (2) Does AI understand what it makes? The first question is now largely settled — AI beats average humans on creativity metrics. The second remains the concept-hard-problem-consciousness applied to machines, and is as open as ever.
Key Facts
- AI vs. 100,000 humans (2026): Researchers at the University of Graz benchmarked leading LLMs (GPT-4, Claude, Gemini) against 100,000+ human participants on the Alternative Uses Task (divergent thinking). Result: AI exceeds average human performance, but the top 10% of human creative thinkers still outperform even the strongest models (ScienceDaily, Jan 2026).
- Individual vs. collective creativity paradox (Science Advances, 2024): AI assistance makes individual stories rated as more creative, better written, and more enjoyable — especially for less creative writers. But AI-assisted stories are more similar to each other than human-only stories. AI boosts individual floors while compressing collective ceilings — diversity of creative output shrinks even as average quality rises.
- Stage-dependent effects (Frontiers in Computer Science, 2025): AI is maximally helpful in the brainstorming stage for everyone. In the implementation stage, it helps non-experts but creates friction for expert designers, who find AI suggestions too generic relative to their refined taste.
- Brain networks of creativity: The Default Mode Network (DMN) is the neural engine of creative thought. A 2025 multi-center study (N=2,433, 10 countries, Nature Communications Biology) found creativity is reliably predicted by the rate of DMN ↔ Executive Control Network (ECN) switching — not sustained DMN activation alone. The creative act requires both spontaneous ideation (DMN) and evaluative control (ECN). Causal stimulation of DMN regions (2024, Brain, Oxford) selectively reduces originality — the closest thing to a causally-confirmed creativity circuit yet.
- The DMN-ECN distance finding (April 2026): Prefrontal cortex acts as the bridge between DMN and ECN. Creativity is predicted by the functional distance between these networks at rest — more distance means less crosstalk, enabling wilder combinations before evaluation. Professional artists show more DMN+ECN+sensorimotor gray matter, less Salience Network expression (bioRxiv, Jan 2025). This inverts a common assumption: creative experts have quieter filtering systems, not louder generative ones.
- Outsider art & FTD paradox: ~2.5% of frontotemporal dementia patients develop de novo visual creativity when frontal inhibition is removed. The same pattern appears in AI: large models generate the most “surprising” combinations precisely because they lack the frontal critic that domain expertise installs. See concept-outsider-art.
What Does “Creative” Mean?
Three definitions in active use:
| Criterion | AI Status |
|---|---|
| Novelty — output not seen before | Achievable; statistical novelty is trivial |
| Value — output is useful or beautiful | Achievable; measurable via human ratings |
| Intentionality — output reflects goals and understanding | Disputed — the Searle objection; see concept-chinese-room |
Margaret Boden’s taxonomy remains the standard framework:
- Combinational creativity — new combinations of existing ideas. AI excels here.
- Exploratory creativity — systematically probing an existing conceptual space. AI is competent.
- Transformational creativity — restructuring the conceptual space itself. No AI system has demonstrably done this; every major paradigm shift in art history (Impressionism, Cubism, Abstract Expressionism) was transformational and remains distinctively human.
The Collective Diversity Problem
The Science Advances 2024 finding is the most important and underreported: AI boosts individual creativity but homogenizes the collective output. The mechanism is statistical: AI models are trained on all existing human work and generate high-probability combinations from that distribution. The tails of human creativity — the outlier, the misfit, the outsider — are precisely the part AI compresses. concept-outsider-art figures like Henry Darger (15,000 hidden pages) and Adolf Wölfli (25,000 psychiatric pages) would be impossible to generate from an AI model trained on the mainstream.
This creates a civilizational risk: if everyone uses the same AI assistants, the genetic diversity of culture narrows. The analog in biology is a concept-svalbard-seed-vault problem — monoculture fragility applied to ideas.
AI and Frisson — Can It Move Us?
concept-frisson (musical chills) arises from prediction violation — the brain expecting one thing and getting another, triggering dopamine. AI-generated music can technically create unexpected harmonic progressions. But the frisson research suggests the listener’s model of the composer’s intention modulates the response. Knowing you’re listening to “an AI that doesn’t understand music” may itself suppress the frisson response. This is unresearched (2026).
Harold Cohen’s AARON — The 50-Year Provocation
Harold Cohen built AARON from 1973–2016, the longest-running AI art system in history. AARON produced thousands of artworks. Cohen insisted to his death that AARON did not understand what it made — but he never resolved whether that mattered. concept-generative-art traces this through from 1965 (Georg Nees) to 2026 (Refik Anadol’s DATALAND, first AI art museum). The question AARON never answered remains the question every LLM inherits: does the system understand what it makes, and if not, does that make it not creative?
The Anthropic Attribution Graph Finding (2025)
Anthropic’s 2025 mechanistic interpretability work (attribution graphs) found that LLMs contain structured semantic intermediate representations — not just statistical pattern matching but genuine concept nodes that causally mediate outputs. concept-chinese-room argued syntax ≠ semantics; the attribution graph result suggests the boundary is blurrier than Searle assumed. The 20% functional self-awareness finding (Oct 2025) deepens this: LLMs have some degree of internal model of their own processing.
None of this proves AI creativity is “genuine” in the phenomenological sense — but it shifts the question from “is there anything there?” to “what exactly is there, and how does it compare to what’s in humans?”
Cross-Realm Connections
- concept-generative-art: traces the history from algorithmic art (1965) to AI art (2026); the conceptual overlap with AI creativity is complete
- concept-outsider-art: the FTD disinhibition paradox and DMN-ECN distance finding both suggest creativity is partly about removing filters, not adding them — the same dynamic that makes AI outputs wide-ranging
- concept-chinese-room: the intentionality question; Anthropic attribution graphs are the most concrete 2025 challenge to Searle’s argument
- concept-frisson: whether AI-generated art can produce genuine aesthetic chills — unresearched but testable
- concept-consciousness and concept-hard-problem-consciousness: the deep question underneath — if there’s no “something it is like” to be the AI, can its outputs be creative in any meaningful sense?
- concept-embodied-cognition: physical intelligence models (2024, $400M) suggest AI creativity may remain constrained until models have grounded sensorimotor experience — creativity as a body-in-world phenomenon